Does AI really understand, or is it just autocomplete?

Short version: a large language model is trained, at its core, to predict the next word, so calling it “advanced autocomplete” isn’t wrong. It’s just not the whole story, in the same way that “a brain is just neurons passing electrical signals” is true but leaves out everything interesting. Very complicated things are often built out of very simple ones, and the history of the universe and of life on Earth is one long example of simple rules producing astonishing results. Researchers looking inside these models have found them planning ahead and combining facts in steps, not just parroting. They also still make confident mistakes, and sometimes explain their reasoning in ways that aren’t what actually happened. Our view: whether or not you call it understanding yet, it’s closer than the word “autocomplete” suggests, and we’re probably only on the cusp.

The case for “just autocomplete”

Let’s be fair to the sceptics first, because they have a real point.

A model like Claude or ChatGPT is trained on an enormous amount of text to do one thing: given everything so far, predict what comes next. Then it’s fine-tuned, with a lot of human feedback, to be helpful and to follow instructions. When you ask it a question, it’s producing an answer one small piece of text at a time, each piece the most fitting continuation of what came before.

The best-known version of the argument is from 2021, when researchers Emily Bender, Timnit Gebru and colleagues described language models as ”stochastic parrots”: systems that stitch together patterns from their training text without any reference to meaning. An older version is the philosopher John Searle’s Chinese room from 1980: imagine a person in a room following a rulebook to answer questions written in Chinese, perfectly, without understanding a word of Chinese. Producing the right output, Searle argued, isn’t the same as understanding.

And there’s evidence on the sceptics’ side. AI still makes things up, fluently and with complete confidence. It can solve a hard problem and then trip over an easy one that any child would get. That’s not how we expect something that “understands” to behave.

Now describe yourself the same way

Here’s the thing that bothers us about “it’s just predicting the next word”. Try the same trick on a human being.

• A brain is just cells sending electrical signals. About 86 billion neurons, each one either firing or not, passing chemical messages across tiny gaps. No single neuron understands anything.
• A neuron is just chemistry. Ions moving across a membrane, proteins folding, molecules binding to other molecules.
• Chemistry is just physics. Electrons and atoms following the same handful of rules everywhere in the universe.

Put like that, the idea that a lump of chemistry could fall in love, write a symphony or understand calculus sounds a bit silly. And yet here we are.

There’s a further twist. One leading theory in neuroscience, called predictive processing, describes the brain itself as a prediction machine: constantly guessing what it will see, hear and feel next, and learning from the gap between the guess and what actually happens. If that’s even partly right, “it just predicts what comes next” may describe us more closely than we’d like.

None of this proves an AI understands anything. What it shows is that describing something by its simplest mechanism doesn’t tell you what it can do. “Just” is doing an enormous amount of work in “just autocomplete”.

Simple rules, astonishing results

Stand back far enough and the whole history of everything is a story of simple things building complicated ones.

The universe is about 13.8 billion years old. It began with no stars, no planets and no chemistry worth mentioning: just energy and the simplest particles, following a small set of physical laws. Those laws, with time, made stars. Stars made the heavier elements. The elements made planets, and on at least one planet, chemistry made life.

Life on Earth goes back at least 3.5 billion years, and for most of that time it was nothing more than single cells. Evolution has no plan and no designer’s understanding. It’s a very simple rule (things that copy themselves slightly imperfectly, and the copies that do better leave more copies) run over and over for an unimaginably long time. That simple rule produced eyes, wings, immune systems and, eventually, brains capable of asking whether a machine can understand.

So if someone says “a system trained only to predict the next word could never really understand”, it’s worth asking: compared to what? Compared to a process that turned single-celled organisms into people with no goal at all, “predict the next word, trillions of times, across most of what humans have ever written” doesn’t sound like such a small thing.

What researchers have found inside the models

The most interesting evidence doesn’t come from what AI says, but from looking inside it while it works. In 2025, Anthropic published research that traced what was happening inside its Claude model as it answered (Tracing the thoughts of a large language model). A few findings stand out:

• It plans ahead. Asked to write a rhyming couplet, the model was expected to pick the rhyming word at the last moment. Instead it chose candidate rhymes before writing the line, then wrote towards them. When the researchers suppressed the word it had planned to use, it switched to a different rhyme and rewrote the line to fit.
• It reasons in steps. Asked for the capital of the state that contains Dallas, it first worked out “Dallas is in Texas”, then “the capital of Texas is Austin”. When the researchers swapped the “Texas” step for “California”, the answer changed to Sacramento. That’s not reciting a memorised answer, it’s combining facts.
• It shares concepts across languages. The same internal features for an idea like “small” or “opposite” lit up whether the question was in English, French or Chinese, with the language only applied at the end. That looks a lot like an idea that exists independently of the words for it.

The same research also found the opposite kind of evidence, which is worth knowing. Sometimes the model’s explanation of its reasoning wasn’t what it actually did: asked to work out a hard maths problem, it described doing a calculation the researchers could find no sign of, and when given a hint it sometimes worked backwards to justify the hinted answer. People do this too, but it’s a good reminder that a fluent explanation isn’t proof of understanding.

Our view: on the cusp

Here’s where we land, and it’s an opinion, not a scientific finding.

Using these tools every day, it doesn’t feel like autocomplete. The best AI models are, frankly, more knowledgeable than anyone I’ve ever met, across more subjects, and they appear to understand what I mean almost perfectly, often better than the people I’m talking to. They pick up on what I’m really asking, notice what I’ve left out, and explain things back in a way that only makes sense if something has grasped the point.

Is that “real” understanding? Honest answer: nobody knows for sure, and part of the problem is that we’ve never had a precise definition of understanding even for ourselves. What we can say is:

• ”Just autocomplete” undersells it. The training is simple. What the training produced isn’t.
• It isn’t human understanding, at least not yet. It doesn’t learn from experience the way you do, it has no body or senses in the usual way, and it can be confidently wrong in ways that a person who understood would spot.
• The direction of travel is clear. These systems have gone from party trick to everyday colleague in a few years. If the history of the universe shows anything, it’s that simple processes, given enough scale and time, can cross lines that once looked impossible.

So we think the honest position is somewhere in between the hype and the dismissal: we may be approaching something like understanding, and we’re probably only on the cusp of it. In ten years, the question “does it really understand?” may sound as dated as “can a computer ever beat a grandmaster at chess?”

What this means day to day

Whatever you decide about the philosophy, a few practical points follow:

• Treat it as capable, not infallible. It can grasp a complicated request better than most people. It can also invent a fact with total confidence. Both are true at once, so check anything that matters.
• Talk to it like a smart colleague. Because it does pick up meaning and context, a clear brief with background gets dramatically better results than a few keywords. (More on that in the slowest part of AI is us.)
• Don’t let the label decide for you. “It’s just autocomplete” is a reason people give for not trying these tools properly. Judge it on what it does with your real work.
• Keep the human judgement. Whether or not it understands, it doesn’t carry responsibility. You do.

Common questions

Sources

All checked 13/10/2026.

• Anthropic, Tracing the thoughts of a large language model (27/03/2025). anthropic.com
• Azevedo et al., Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain, Journal of Comparative Neurology (2009): about 86 billion neurons. pubmed.ncbi.nlm.nih.gov
• European Space Agency, Planck reveals an almost perfect Universe (age of the universe, 13.8 billion years). esa.int
• Stanford Encyclopedia of Philosophy, The Chinese Room Argument. plato.stanford.edu
• Bender, Gebru, McMillan-Major and Shmitchell, On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, FAccT 2021. dl.acm.org